Content
75%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A tight, instruction-only skill body that gives specific, operational rules and guardrails rather than vague direction, with a clear decision-driven workflow and validation checkpoints. The 6-signal table and output spec are highly actionable; adding one concrete before→after example would round it out.
Suggestions
Add one short before→after example pair inline (e.g. for C-8 대구) so the prescription is unambiguous without leaving the body.
Tighten the '이건 일부러 안 본다' section to a compact list; the rationale for each omission can be one phrase rather than a clause.
Make the 30% changed-ratio guard an explicit validate→rollback feedback step in the workflow section so the checkpoint is named as a loop.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Dense and operational with almost no padding of concepts Claude already knows; the '이건 일부러 안 본다' justification section and some guardrail elaboration could be trimmed slightly, but every section earns its place. | 4 / 5 |
Actionability | Concrete prescriptions per signal ('2회+', '쉼표 삭제', '종결어미만 변주'), a specified 4-part output format, and a 3-signal decision threshold; minor gap is the absence of an illustrative before→after example in the body itself. | 4 / 5 |
Workflow Clarity | Clear sequence (input → scan 6 signals → count triggers → branched output) with validation checkpoints — '안 고쳐 보고 숫자를 쓰지 않는다', the 30% changed-ratio guard, and the rollback rule — leaving only minor explicit feedback-loop gaps. | 4 / 5 |
Progressive Disclosure | Single well-sectioned file with clear headers; the one external reference (humanize-korean/references/empirical-validation.md) is clearly signaled as the evidence basis and is one level deep, with only minor organization room remaining. | 4 / 5 |
Total | 16 / 20 Passed |